A nearest neighbour clustering approach for incipient fault diagnosis of power transformers

Abstract

Dissolved gas analysis (DGA) is one of the popular and widely accepted methods for fault diagnosis in power transformers. This paper presents a novel DGA technique to improve the diagnosis accuracy of transformers by analysing the concentrations of five key gases produced in transformers. The proposed approach uses a clustering and cumulative voting technique to resolve the conflicts and deal with the cases that cannot be classified using Duval Triangles, Rogers’ Ratios and IEC Ratios Methods. Clustering techniques group the highly similar faults into a cluster providing a virtual boundary between dissimilar data. A cluster of data points may contain single or multiple types of faulty transformers’ data with different distinguishable percentages. The k-nearest neighbour (KNN) algorithm is used for indexing the three closest clusters from an unknown transformer data point and allows them to vote for single or multiple faults categories. The cumulative votes have been used to identify a transformer’s fault category. Performance of the proposed method has been compared with different conventional methods currently used such as Duval Triangles, Rogers’ Ratios and IEC Ratios Method along with published results using computational and machine learning techniques such as rough sets analysis, neural networks (NNs), support vector machines (SVMs), extreme learning machines (ELM) and fuzzy logic. The experimental comparison with both published and utility provided data show that the proposed method can significantly improve the incipient fault diagnosis accuracy in power transformers.